Machine learning taxonomy-based fragility derivation for bridge portfolios under exogenous hazards: application to simply supported prestressed concrete bridges

This paper presents a Machine Learning taxonomy-based framework for deriving fragility curves of bridge portfolios under exogenous hazards, with application to multi-span simply supported prestressed concrete girder bridges with single-shaft reinforced concrete piers. A tailored taxonomy is created to account for two main hazards: traffic and seismic loads. Different finite element numerical models are created for substructure and superstructure, aimed at performing parametric analyses accounting for material degradation (i.e., corrosion). The results are used as input dataset to train and test Machine Learning algorithms to derive fragility curves for different collapse limit states. Statistical metrics and eXplainability approaches are used to identify the best surrogate model and the most critical input features. The comparison with finite element-based fragility curves shows median differences generally lower than 5% for traffic hazard and lower than 10% for seismic hazard. The proposed approach is tested on a real-life case study to assess the accuracy of prediction against numerical analyses for rapid portfolio-scale screening and prioritization. The main innovation of the work is a tangible and quantitative set of Machine Learning taxonomy-based fragility curves, which can be used for supporting prioritization of bridge portfolios.